TDA4VM: UINT8 tidl output dequantize to float output

Ma Qiang

Part Number: TDA4VM

Recently, we get some problems about uint8 model output dequantize to float. After tidl model importing, we can get a debug info of each layer, including scale, meandiff and so on. We use the scale and meandiff to trans our uint8 output to float32, but cannot get the right output.

So, we just want to know if there are some special ops to dequantizing the uint8 output.

Regards

QIang

over 3 years ago

+1 Anshu Jain over 3 years ago

TI__Guru 56820 points

Hi QIang,

Please refer following document to understand how to de-quantize TIDL output :

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_00_00_12/exports/docs/tidl_j7_08_00_00_10/ti_dl/docs/user_guide_html/md_tidl_fsg_io_tensors_format.html#did_tidl_io_format_Types

Regards,

Anshu

0 Ma Qiang over 3 years ago in reply to Anshu Jain

Prodigy 110 points

Hi Anshu

Thanks, i will close this thread

Regards,

QIang

0 Ma Qiang over 3 years ago in reply to Anshu Jain

Prodigy 110 points

can you tell me how can i get the outTensorScale[]. We use rtos7.3 ,there is no output tensor scale in struct sTIDL_IOBufDesc_t.

Regards

Qiang

0 Anshu Jain over 3 years ago in reply to Ma Qiang

TI__Guru 56820 points

Hi Qiang,

This parameter was added from SDK 8.0 onwards. Are you trying to read this parameter in Host/PC emulation mode? If yes then you can read the same from TIDL_outArgs->scale[]

Regards,

Anshu

+1 Ma Qiang over 3 years ago in reply to Anshu Jain

Prodigy 110 points

Hi Anshu

We have add our post proc kernel in tda4vm and compare the results with pc post on target c66. And just confuse why the output was different. We doubt the problems was caused by dequantize.

I will try to debug TIDL_OutArgs, thanks.

Regards

Qiang

0 Anshu Jain over 3 years ago in reply to Ma Qiang

TI__Guru 56820 points

Hi Qiang,

Can we close this thread? You can create a separate thread if you face further issues?

Regards,

Anshu

Processors

Processors forum

TDA4VM: UINT8 tidl output dequantize to float output